Comparing visual spec-to-app model scores over time (updated periodically)
| Rank | Model | Score | Delta | Notes |
|---|---|---|---|---|
| 1 | GPT-4o (2024-08) | 92 | +4 | Strong spec compliance, handles complex UIs |
| 2 | Claude 3.5 Sonnet | 88 | +2 | Excellent layout accuracy, slower generation |
| 3 | Gemini 1.5 Pro | 85 | +3 | Good for simple forms, struggles with animations |
| 4 | Claude 3 Opus | 83 | -1 | Baseline benchmark from Q2 2024 |
| 5 | GPT-4 Turbo | 80 | -2 | Still solid for basic spec-to-app tasks |
| 6 | Llama 3.1 405B | 74 | +5 | Open-source leader, improving rapidly |
| 7 | Mistral Large 2 | 70 | +1 | Good for minimal UIs, lacks image support |
| 8 | Gemma 2 27B | 65 | +3 | Promising small model for mobile UIs |
Copy the link to share this benchmark board:
The source hardcodes eight leaderboard rows but plots four series over six quarters. The heading says last four quarters, yet the actual axis runs from first-quarter twenty twenty-four through second-quarter twenty twenty-five. Bars at forty display pixels per count measure three hundred twenty for eight rows, one hundred sixty for four series and two hundred forty for six quarters. There is no benchmark dataset, test procedure, live updating or score verification in this source. The model names and performance notes are supplied fixture content, not evidence of current model performance. The chart y axis spans forty to one hundred. Its fixed last values are ninety-two, eighty-eight, eighty-five and seventy-four. Above a shared forty-point baseline those spans are fifty-two, forty-eight, forty-five and thirty-four points. At five display pixels per point, corresponding bars are two hundred sixty, two hundred forty, two hundred twenty-five and one hundred seventy. The difference between ninety-two and eighty-eight is four points, not four percent relative improvement. Straight numerical readings matter more than a visual impression from the truncated axis. These are source constants only. All four plotted series repeat their fifth value at the sixth quarter, so their last adjacent change is zero. For the first row the table nevertheless shows plus four. Its sequence seventy, seventy-five, eighty-two, eighty-eight, ninety-two, ninety-two gives consecutive differences five, seven, six, four and zero. Difference bars at thirty pixels per point have widths one hundred fifty, two hundred ten, one hundred eighty, one hundred twenty and zero. The plus-four badge matches the earlier eighty-eight-to-ninety-two step, not the last repeated quarter. Legend clicks only hide or show source chart series. Copy Link copies a placeholder URL rather than creating a hosted benchmark.